Accurate forecasting of gold prices, due to their nonlinear and non-stationary nature and their susceptibility to numerous economic and financial factors, is considered one of the complex problems in financial time-series analysis. The aim of this study is to design and evaluate an accurate, stable, and interpretable model for short-term forecasting of global gold prices using the Temporal Fusion Transformer (TFT) architecture.
The present study is applied in terms of purpose and quantitative, data-driven, and based on computational experiments in terms of methodology. The data used include daily gold market information over a ten-year period from 2014 to 2024. After data cleaning, normalization, and preprocessing, the dataset was temporally divided into training, validation, and test sets. The proposed model was implemented to forecast gold prices over 1-, 5-, 10-, and 22-day horizons, using variable selection networks, gated linear units, recurrent layers, and an interpretable multi-head attention mechanism. The performance of the model was compared with ARIMA, SARIMA, LSTM, GRU, standard Transformer, and Informer methods.
The results showed that the TFT model achieved the best performance among the compared methods, with an RMSE of USD 18.42 per ounce, MAE of USD 13.87 per ounce, MAPE of 0.89%, and a coefficient of determination (R²) of 0.9534. Furthermore, the model achieved a directional accuracy of 72.5% and a prediction interval coverage of 79.8%. The Diebold–Mariano test also confirmed the statistically significant superiority of the proposed model over most benchmark methods. The findings indicate that, in addition to improving forecasting accuracy, the TFT model has an appropriate capability to provide prediction intervals and interpret the importance of influential variables and time periods. Therefore, it can be used as an effective tool for decision-making by investors and gold market analysts.